DirectX 8 Graphics Overview

Dave Horne
Technical Developer Relations
NVIDIA Corporation
 Initialisation Changes

• Merged DirectDraw and Direct3D

• Simplified API
• Improved data allocation and performance
• Promoting overall robustness
• App code simplification
• Fewer test paths

NVIDIA Proprietary and Confidential

 Simplified API

• Don't need to use QueryInterfaces, and no GUIDs

in DirectX 8
• One line global D3D object creation
• m_pD3D = Direct3DCreate8(D3D_SDK_VERSION);
• D3D_SDK_VERSION ensures compilation with
correct headers
• No more callbacks!
• Query Adapters(physical pieces of hardware),
devices and modes directly from D3D object
• GetDeviceCaps() replaces GetCaps()
• called from D3D object
• No need to create devices to determine caps

NVIDIA Proprietary and Confidential

 Device Types

• HAL device
• Supports hardware accelerated rasterization and
both hardware and software vertex processing
• Reference device
• Supports all features of API correctly (if not
quickly!) in software only. Intended for feature
testing, demos, etc, not for retail apps
• Software device
• DX8 supports ‘pluggable’ software devices (DDK)
• MS will not be providing a SW device
• HAL or REF device variant : Pure device
• Hardware only vertex processing
• Highly restricted querying of device state
• Performance advantage over non-pure devices
NVIDIA Proprietary and Confidential
 Device Creation I

• Set ‘Presentation Parameters’ structure

• Creation of back buffer(s)
• Width, height and bit depth for full-screen
• Windowed/Full-screen
• Specifying buffer ‘Swap’ type
• Flip or Copy(Blit)
• Choose auto depth/stencil creation
• Multi-sample type
• Anti-aliasing, etc
• Full-screen refresh rate
• Can set 0 for some parameters to get default
• Set D3DPRESENTFLAG_LOCKABLE_BACKBUFFER if
you need access to the back buffer
• Setting this flag will reduce performance
NVIDIA Proprietary and Confidential
 Device Creation II

• Presentation parameters structure very similar

for windowed and full-screen device creation
• Also used in device Reset() from lost device
• Behaviour flags in CreateDevice()
• D3DCREATE_*_VERTEXPROCESSING
• Software, Hardware or Mixed
• D3DCREATE_FPU_PRESERVE
• D3DCREATE_MULTITHREADED
• D3DCREATE_PUREDEVICE
• Limited renderstate shadowing
• Hardware vertex processing only

NVIDIA Proprietary and Confidential

 Device Creation III
D3DPRESENT_PARAMETERS d3dpp;
memset(&d3dpp, 0, sizeof(d3dpp));

d3dpp.BackBufferWidth = 1024; // must be 0 in windowed mode

d3dpp.BackBufferHeight = 768; // must be 0 in windowed mode
d3dpp.BackBufferFormat = D3DFMT_A8R8G8B8;
d3dpp.BackBufferCount = 1;
d3dpp.Windowed = FALSE;
d3dpp.SwapEffect = D3DSWAPEFFECT_FLIP;
d3dpp.EnableAutoDepthStencil = TRUE;
d3dpp.AutoDepthStencilFormat = D3DFMT_D32;
d3dpp.MultiSampleType = D3DMULTISAMPLE_NONE;
d3dpp.hDeviceWindow = g_hwnd;

/* Following elements must be zero for Windowed mode */

d3dpp.FullScreen_RefreshRateInHz = D3DPRESENT_RATE_DEFAULT;
d3dpp.FullScreen_PresentationInterval = D3DPRESENT_INTERVAL_DEFAULT;

pEnum->CreateDevice(
D3DADAPTER_DEFAULT, // iAdapter
D3DDEVTYPE_HAL, // device type
hWindow, // focus
D3DCREATE_FPU_PRESERVE, // dwBehaviorFlags
&d3dpp, // presentation parameters
&pDevice); // [out] pDevice

NVIDIA Proprietary and Confidential

 Lost Devices I

• Supercedes DX7 Lost Surfaces

• Device can be in an operational or a lost state
• Lost state caused by loss of focus, Alt-Tab, power
management events, etc.
• Once in lost state only error returns are from
• Present()
• TestCooperativeLevel()
• All other methods succeed
• Even if they don’t do anything

NVIDIA Proprietary and Confidential

 Lost Devices II

• App needs to detect and respond to Lost device

• Rendering surfaces are gone
• All unmanaged video/agp mem resources are invalid
• includes VBs and Index buffers as well as textures
• All HW states, lights and shaders are invalid
• Rebuild the device using the Reset() command
• App MUST free all unmanaged video/agp resources
• Use same presentation params as CreateDevice()

• App must recreate any unmanaged resources, as

well as renderstates, lights and shaders

NVIDIA Proprietary and Confidential

 Lost Devices III
// Test the cooperative level to see if it's okay to render
// Call once per render loop
if(FAILED(hr = m_pD3DDevice->TestCooperativeLevel()))
{
// If the device was lost, do not render until we get it back
if(D3DERR_DEVICELOST == hr) return S_OK;
// Check if the device is ready to be reset
if(D3DERR_DEVICENOTRESET == hr)
{
// Call app specific non-d3d managed resource destruction
DestroyAppManagedResources();
// Resize the device
if(FAILED(hr = m_pD3DDevice->Reset(&m_d3dpp))) return hr;
// Call app specific non-d3d managed resource creation
CreateAppManagedResources();
// Restore renderstates, lights, shaders, etc.
RestoreOtherStuff();
}
return hr;
}

NVIDIA Proprietary and Confidential

 Resources
• Unified resource manager
• Geometry
• Vertex buffers
• Index buffers
• Textures
• 2D Textures
• Volume textures
• Cube maps (CubeTexture)

NVIDIA Proprietary and Confidential

 Resource Creation I

• Explicit types
• CreateVolumeTexture() distinct from
CreateCubeTexture(), CreateTexture()
• No CreateSurface(),GetAttachedSurface()
• Pool
• D3DPOOL_DEFAULT
• Placed in memory type based on device and usage
• Always lost on device loss
• D3DPOOL_SYSTEMMEM
• Placed in sys mem.
• Never lost
• D3DPOOL_MANAGED
• Managed by D3D or driver.
• Never lost
• Can't use with D3DUSAGE_DYNAMIC flag
NVIDIA Proprietary and Confidential
 Resource Creation II

• Usage
• D3DUSAGE_DEPTHSTENCIL, D3DUSAGE_RENDERTARGET
• Indicates using as depth/stencil or render target
• D3DUSAGE_SOFTWAREPROCESSING
• D3DUSAGE_DONOTCLIP
• D3DUSAGE_DYNAMIC, D3DUSAGE_WRITEONLY
• Allows driver to place resource in optimal location
• VB only ‘hints’ to allow the driver to put VB in CPU
accessible memory if no native HW support
• D3DUSAGE_POINTS
• Point sprites
• D3DUSAGE_RTPATCHES
• 'RT' stands for rectangular and triangular high order surfaces
• D3DUSAGE_NPATCHES

NVIDIA Proprietary and Confidential

 Resource Creation III

• Format
• Now an enumerated type
• Runtime predefines several types of each class
• Color, luminance, DXTn, vertex, index, etc
• IHVs may still add vendor specific formats via the
FOURCC codes
• Always allocate all D3DPOOL_DEFAULT resources
before D3DPOOL_MANAGED
• Otherwise D3D resource manager will get an
inaccurate picture of available memory

NVIDIA Proprietary and Confidential

 DX8 Vertex Buffer Changes I

• Vertex buffers are now the standard usage for

DrawPrimitive()and DrawIndexedPrimitive()
• ‘User Pointer’ calls for apps that can’t use VBs for
some reason
• DrawPrimitiveUP()
• DrawIndexedPrimitiveUP()
• Don’t use! Will almost certainly result in a redundant
copy of vertex data by driver
• VBs can consist of arbitrary data in multiple streams
• Use Vertex Shaders to interpret data

NVIDIA Proprietary and Confidential

 DX8 Vertex Buffer Changes II

• CreateVertexBuffer() changes
• Size now defined as number of bytes
• DX7 defined it as number of vertices
• Usage flags
• FVF code can be zero
• Called a Non-FVF VB
• Represents VB stream for vertex shader
• Format implied by CreateVertexShader() declaration
• Must declare which memory pool required
• No more maximum VB size
• DX7 max size is 0xffff

NVIDIA Proprietary and Confidential

 Index Buffers I

• New resource type for DX8

• Removes redundant driver copy of indices in the
same way VBs stop driver copy of vertices
• CreateIndexBuffer() size, usage and pool
parameters the same as VB create
• Format parameter allows creation of either 16 or 32
bit indices
• D3DFMT_INDEX16, D3DFMT_INDEX32
• 32 bit indices new to DX8

NVIDIA Proprietary and Confidential

 Index Buffers II

• Semantics the same as for VBs

• Create
• Lock()
• Fill
• Unlock()
• Use
• Indices now explicitly set via SetIndices()
• In DX7 they were part of the DP/DIP calls
• 'User Pointer' calls still require explicit index
parameters
• Similar to DX7
• Currently set IB is ignored
• Reminder: Don't use User Pointers!

NVIDIA Proprietary and Confidential

 Locking Resources I

• Grants CPU access to resource

• Texture locks
• LockRect() for 2D textures and cube maps
• LockBox() for volume textures
• Typically only one lock per resource
• Geometry locks
• Lock() for both index and vertex buffers
• Multiple locks per buffer allowed
• Must relinquish lock with UnLock() before
resource can be used by device
• Any device operations with locked resource will
be serialised until resource unlocked
NVIDIA Proprietary and Confidential
 Locking Resources II

• General lock flags

• D3DLOCK_READONLY
• Won’t write to resource so no recompression on unlock
• Can’t use with VBs/IBs created with WRITEONLY flag
• D3DLOCK_NOSYSLOCK
• No system-wide critical section taken
• Texture specific lock flags
• D3DLOCK_NO_DIRTY_UPDATE
• Don’t update dirty region
• Geometry specific lock flags
• D3DLOCK_DISCARD
• D3DLOCK_NOOVERWRITE

NVIDIA Proprietary and Confidential

 Locking Textures I

• LockRect() can lock whole surface or sub-rect

of 2D texture or face of cube map
• Lock with RECT or NULL for entire surface
• Returns D3DLOCKED_RECT
• Consists of data pointer and pitch (in bytes)
• Simpler than DX7 surface lock return
DDSURFACEDESC2
• Must specify mip level required for lock
• Cube textures must also specify face

NVIDIA Proprietary and Confidential

 Locking Textures II

• Use LockBox() for to lock volume or sub-

volume of 3D texture
• Lock with D3DBOX structure or NULL for entire
volume
• Must also specify mip level
• All 3 dimensions of each level are divided by 2
(rounding down) down to minimum 1x1x1
• Returns D3DLOCKED_BOX
• Consists of data pointer, row pitch and slice pitch
• Compressed DXT formats
• Can only be locked on 4x4 boundaries
• Minimum actual size is 4x4 on a side
• Textures or mip levels less than this are padded
NVIDIA Proprietary and Confidential
 Locking Geometry

• DX8 Lock() on VBs and IBs allows sub-ranges to

be specified
• DX7 could only lock whole VB
• Multiple locks on single buffer permitted
• Allows efficient usage of large VBs containing
multiple models, as you can render from one sub
range whilst locking and modifying another
• D3DLOCK_DISCARD and D3DLOCK_NOOVERWRITE
flags are valid only on buffers created with
D3DUSAGE_DYNAMIC
• Therefore can't be managed buffers

NVIDIA Proprietary and Confidential

 Programmable Vertex Shaders

• New feature for DX8

• Can be used in place of D3D 'fixed function' T&L
pipeline
• Allows extremely versatile geometry handling by
using vertex programs to process vertex data
• Operates at the vertex level only
• No knowledge of other vertices (no primitives)
• Cannot create or destroy vertices
• Vertex output must be in homogenous clip-space
with colour, texture coordinate, fog, and sprite
point size as needed
NVIDIA Proprietary and Confidential
 Streams

• New for DX8

• Allows multiple VBs to be set as data sources for
vertex processing
• Stream elements are combined at draw time
• Single set of indices for all streams in DX8
• GeForce/GeForce2 can support up to 8 vertex
streams, fed to the fixed-function T&L pipeline
only
• DX8 level hardware must be able to support up to
16 vertex streams, fed to the fixed-function OR
vertex shading pipeline
NVIDIA Proprietary and Confidential
 Creating Vertex Shaders

• Use CreateVertexShader()
• Pass in a declaration defining input to shader
• Define format of vertex streams and bindings to vertex
shader input registers
• Bind primitive tessellator to input registers
• Loading of data into constant memory
• Pass in pointer to vertex shader program token data
• E.g. from D3DXAssembleShader
• Or use NULL to get fixed function processing
• Usage flag
• D3DUSAGE_SOFTWAREPROCESSING if you're using
software vertex processing
• Returns handle for use in SetVertexShader()

NVIDIA Proprietary and Confidential

 Setting Streams

• Call SetStreamSource()
• Stream index
• 0 to (maximum number of streams - 1)
• Pointer to VB
• Stride of vertex
• For FVF VBs must be the size of vertex computed
from FVF code
• For non-FVF VBs must match the size implied in the
CreateVertexShader() declaration
• Number and format of streams set must match
the current vertex shader declaration

NVIDIA Proprietary and Confidential

 Using Vertex Shaders

• Ensure that all VB streams have been set

• Set an IB with SetIndices() if using indexed data
• Use SetVertexShader() to select shader
• Pass in a handle from CreateVertexShader()
• Or pass a valid FVF code to use the fixed function
pipeline
• Use DrawPrimitive() or DrawIndexedPrimitive()
to render primitives

NVIDIA Proprietary and Confidential

 Fixed Function Pipeline

• Can use the existing 'fixed function' T&L pipeline

with standard flexible vertex format codes
• Use SetVertexShader() with an FVF code instead
of a vertex shader handle
• Only stream 0 is valid
• Example pseudocode
gVB = CreateVertexBuffer(); // with FVF code
// Lock, fill, unlock gVB
SetVertexShader(); // with same FVF code
SetStreamSource(0, gVB, sizeof(Vertex));
DrawPrimitive();
NVIDIA Proprietary and Confidential
 Fixed Function - Multiple Streams
• Create and set a vertex shader with declarator
describing streams, and NULL program
• Create VBs for each of the separate streams, and set
them as stream sources
• Pseudocode for a 2 stream XYZ and colour multiple
stream fixed function shader
hVS = CreateVertexShader(pDec, NULL,...);
gVB_Pos = CreateVertexBuffer(FVFXYZ);
gVB_Col = CreateVertexBuffer(FVFCOLOR);
SetVertexShader(hVS);
SetStreamSource(0, gVB_Pos, sizeof(XYZ));
SetStreamSource(1, gVB_Col, sizeof(COLOR));
DrawPrimitive();

NVIDIA Proprietary and Confidential

 Scene Presentation

• New device function Present()

• Subsumes DX7 Flip() and Blit()
• Consistent for full-screen and windowed
• Simplifies renderloop
• No more windowed app lag due to excessive blit
buffering
• Swap chain now a property of device
• Number of buffers specified in device creation
• Always at least one chain
• Additional ones can be created if needed
• Front buffer now part of device
NVIDIA Proprietary and Confidential
 Front Buffer

• No more concept of ‘Primary Surface’

• Front buffer not directly accessible
• Cannot lock or render to
• Mouse cursor API available
• Hardware if available
• Can get to front buffer via GetFrontBuffer()
• Guaranteed to be slow
• Only really useful for debugging

NVIDIA Proprietary and Confidential

 Depth/Stencil Surfaces

• Z buffer now set into the device

• AutoDepthStencil
• Set in presentation params – created, destroyed and
resized automatically by the device
• Explicit DepthStencil creation
• Created by app, and attached to the device
• Use SetRenderTarget()
• Useful when switching render targets halfway
through a scene
• No auto resize on mode or window size changes
• No auto destruction

NVIDIA Proprietary and Confidential

 Reset() (again)

• As well as responding to lost devices, Reset()

can be used to
• Resize the buffers
• Change buffer format
• Swap between full-screen and windowed
• Switch stereo on and off
• Calling Reset() causes all unmanaged video/agp
memory resources, and all state information, to
be lost.
• So similar (if not identical) code path to recovery
from device loss

NVIDIA Proprietary and Confidential

 Summary

• DX8 initialisation is far easier than DX7

• Device creation and loss management improved
• New Index Buffer resource gives VB style
management for indices
• Vertex streams allows vertex components in
separate VBs
• Vertex shaders will give unprecedented
programmability in the transform pipeline
• Scene presentation simplified

NVIDIA Proprietary and Confidential